Fair and trustworthy biometric systems: From robustness analysis to fairness-aware learning

Biometric systems based on artificial intelligence have reached a level of performance that enables their large-scale deployment in operational and societal contexts. At the same time, concerns have emerged regarding their robustness, fairness, and overall trustworthiness. In particular, systematic performance disparities across population groups and sensitivity to real-world data perturbations raise questions about the reliability and equity of such systems.
This thesis investigates trustworthiness in biometric systems through an evaluation-driven perspective grounded in international standards. Biometric recognition performance is consistently analyzed using the False Match Rate (FMR) and the False Non-Match Rate (FNMR), as defined in ISO/IEC 19795-1:2021, and studied under realistic acquisition conditions, demographic variations, and intentional perturbations.
The first part of the thesis focuses on robustness and analyzes the impact of social media processing, face beautification filters, image and light-field compression, and adversarial perturbations on biometric systems. The results show that visually benign transformations can induce significant variations in biometric error rates, revealing vulnerabilities that are not captured by conventional image-quality measures.
Building on these findings, the thesis addresses fairness through the study of demographic information encoded in biometric representations. Through the analysis of foundation-model face embeddings based on Vision Transformers, it is shown that a single representation can support the estimation of multiple demographic and physiological attributes. Motivated by this observation, the thesis investigates demographic performance disparities in biometric systems through the evaluation of fairness metrics for automatic speaker verification, the analysis of fairness– privacy–utility trade-offs, and the development of fairness-aware representation learning strategies designed to mitigate shortcut learning and feature entanglement.
Taken together, these contributions establish a methodological foundation for the development and evaluation of fair and trustworthy biometric systems. By linking robustness analysis, representation learning, and fairness assessment, this thesis contributes to the understanding of bias in biometric technologies and supports their responsible deployment in real-world applications.